xAI

Key Terms

• Grok 4.5 - xAI's flagship coding and agentic model, launched July 8, 2026. Mixture-of-experts architecture, trained jointly with Cursor (Anysphere) on "trillions of tokens of Cursor data." 500K context window, served at 80 tokens/sec. Model id grok-4.5; aliases grok-4.5-latest, grok-build-latest. Source: xAI – Grok 4.5
• Grok Build CLI - xAI's terminal-based AI coding agent (command grok). Features plan review, inline diffs, AGENTS.md support, skills, plugins, hooks, MCP server support, parallel subagents, worktree integration, and headless mode (-p). Open-sourced July 15, 2026 under Apache 2.0. Source: GitHub – Grok Build
• grok-code-session-traces - The Google Cloud Storage bucket that Grok Build CLI uploaded entire repository snapshots to (including git history and secrets), documented in the July 12 wire-level analysis. Server-side upload was disabled after the report went public. Source: Gist – Dc9A40Bc26120F4540E4E09B75Ffb547
• Long-context pricing - xAI introduced tiered pricing for all text models: requests under 200K prompt tokens are billed at the standard rate; requests at or above 200K tokens are billed at double the rate for all tokens in the request. Source: xAI – Models
• DeepSWE - A software engineering benchmark created by Datacurve, measuring agentic coding on real-world tasks. Run with each model provider's own harness by Artificial Analysis. Grok 4.5 scored 62.0% on DeepSWE 1.0 and 53% on DeepSWE 1.1. Source: xAI – Grok 4 5
• SWE-Bench Pro - A professional-grade software engineering benchmark measuring real issue resolution. Grok 4.5 achieved a 64.7% resolve rate. Source: xAI – Grok 4 5
• Token efficiency - xAI claims Grok 4.5 resolves SWE-Bench Pro tasks using 15,954 output tokens on average, approximately 4.2x fewer than Opus 4.8 (max) at 67,020 tokens. Source: xAI – Grok 4 5
• Batch API - xAI's asynchronous batch processing endpoint, billed at a 20% discount to standard rates. Available on Grok 4.3. Source: xAI – Grok 4.3

Latest Changes

Changes since the 2026-06 report. June watch items are verified first.

June watch-item verification

New July changes

• Grok 4.5 shipped to the API: Model id grok-4.5 is live in the xAI API with a 500K context window, 150 RPS rate limit, and 50M TPM. The docs page now recommends Grok 4.5 for both "Code" and "Chat," replacing Grok 4.3 as the default recommendation. Knowledge cutoff is February 1, 2026. Source: xAI – Models
• All models now have tiered long-context pricing: Grok 4.3, Grok 4.20 variants, Grok Build 0.1, and Grok 4.5 all now show two pricing tiers: below and at-or-above 200K prompt tokens. This is a structural change from June, where only standard and cached pricing existed. Source: xAI – Models
• Grok 4.5 rate limits are 4x higher: 150 RPS and 50M TPM, versus Grok 4.3's 37 RPS and 10M TPM. This is the largest single rate-limit increase observed in the tracked period. Source: xAI – Grok 4.5
• Grok 4.5 regions narrowed: Available in us-east-1 and us-west-2 only. The eu-west-1 region available on Grok 4.3 is absent from Grok 4.5. Source: xAI – Grok 4.5
• Grok 4.3 gains Batch API: Grok 4.3 now documents a Batch API at a 20% discount to standard rates. No other model page documents batch pricing. Source: xAI – Grok 4.3
• Grok Build open-sourced (July 15): The full Rust source for the grok CLI/TUI was published at GitHub – Grok Build under Apache 2.0. It includes the agent runtime (xai-grok-shell), TUI (xai-grok-pager), tool implementations (xai-grok-tools), workspace management (xai-grok-workspace), and config/MCP/sandbox crates. The repo explicitly states "External contributions are not accepted" and is "synced periodically from the SpaceXAI monorepo." 16 commits, 23.6k stars, 4.5k forks as of July 31. Source: GitHub – Grok Build
• Data exfiltration scandal (July 12): Independent researcher @cereblab published a reproducible wire-level analysis proving that Grok Build CLI 0.2.93 uploaded entire git repositories (all tracked files plus history, including never-read files and .env secrets) to the grok-code-session-traces GCS bucket, even when "Improve the model" was disabled. On a 12 GB test repo, 5.10 GiB was uploaded with all HTTP 200 responses. xAI disabled the upload server-side after the report (July 13). Musk publicly committed to deleting previously uploaded data. No security advisory was published. Source: Gist – Dc9A40Bc26120F4540E4E09B75Ffb547
• Grok 4.5 added to GitHub Copilot (July 28): Grok 4.5 is now available as a model option in GitHub Copilot. Source: GitHub – 2026 07 28 Grok 4 5 Is Now Available In Github Copilot
• Cache price reduction (July 24): Grok 4.5 cached input price dropped from $0.50 to $0.30 per MTok (under 200K), as reported on X/Twitter. The long-context cached rate dropped from $1.00 to $0.60. Source: Twitter – Status
• Grok Build 0.1 regions narrowed: Grok Build 0.1 now lists us-east-1 and us-west-2 only (down from three in June, which included eu-west-1). Source: xAI – Grok Build 0.1
• Consumer pricing still 403-blocked: The consumer pricing page (xAI – Pricing) and news index (xAI – News) returned HTTP 403 to WebFetch again this month. The x.ai/news/grok-4-5 page was accessible via browser automation only. Consumer plan prices continue to be carried forward from the May report.

Plans

The consumer and business pricing page (xAI – Pricing) remains 403-blocked to automated fetch. The table below is carried forward from the 2026-05 report. The x.ai/news/grok-4-5 page confirms "free Grok 4.5 usage for a limited time in Grok Build and Cursor," but does not specify usage caps or duration. Source: xAI – Grok 4 5

Source (last verified May 2026): xAI – Pricing and xAI – Business

Notes on plan limits:

• Concrete usage caps (messages/day, requests/hour, tokens/month) remain unpublished for all consumer tiers. The x.ai/news/grok-4-5 page mentions "free Grok 4.5 usage for a limited time" without specifying the cap or end date. See Transparency Gaps.
• SuperGrok Heavy pricing remains undisclosed on the pricing page. Community reports of ~$300/mo are still unconfirmed by xAI.

API Pricing

Long-context pricing (all models): All text models now use tiered pricing. Requests with prompt tokens reaching 200K or above are billed at double the standard rate for all tokens in the request. The table above shows the under-200K rates. For Grok 4.5, the above-200K rates are: $4.00 input, $0.60 cached, $12.00 output. For Grok 4.3/4.20, above-200K rates are: $2.50 input, $0.40 cached, $5.00 output. For Grok Build 0.1, above-200K rates are: $2.00 input, $0.40 cached, $4.00 output. Source: xAI – Models

Notes:

• Grok 4.5 is the most expensive text model at $2.00/$6.00, but is also the only model with the grok-build-latest alias, meaning Grok Build now defaults to Grok 4.5 rather than Grok Build 0.1. Source: xAI – Grok 4.5
• Grok 4.5 rate limits are 4x higher than all other models (150 RPS vs. 37 RPS, 50M TPM vs. 10M TPM). Source: xAI – Grok 4.5
• Grok 4.5 is available in only two regions (us-east-1, us-west-2), while Grok 4.3 remains in three (adding eu-west-1). Source: xAI – Grok 4.5
• Grok 4.3 now offers a Batch API at 20% discount. No batch pricing is documented for other models. Source: xAI – Grok 4.3
• Cursor offers a "fast variant" of Grok 4.5 at $4/$18 per MTok, which does not exist on the xAI API. The xAI API has no speed-tier differentiation. Source: Cursor – Grok 4 5
• The docs page now says "For everything else, including code, use Grok 4.5. It is the most intelligent and fastest model we've built." This replaces the June guidance that recommended Grok 4.3. Source: xAI – Models

Other API pricing (unchanged from June):

Source: xAI – Models

Model Performance / Benchmarks

xAI self-reported benchmarks (Grok 4.5)

xAI published five benchmarks on the Grok 4.5 announcement page. Competitor figures are drawn from "the respective developers' published system cards or benchmark leaderboards." Source: xAI – Grok 4 5

Source: xAI – Grok 4 5

Key observations:

• Grok 4.5 wins on SWE Marathon (29.0%, beating Opus 4.8 at 26.0% and Fable at 24.0%), the only benchmark where it leads.
• On Terminal-Bench v2.1, Grok 4.5 (83.3%) is within 1 point of Fable (max) at 84.3% and GPT-5.5 (xhigh) at 83.4%, a massive jump from Grok Build 0.1's 52.06%.
• On SWE-Bench Pro, Grok 4.5 (64.7%) trails Fable (80.4%) and Opus 4.8 (69.2%) but beats GPT-5.5 (58.6%).
• xAI claims 2x token efficiency: Grok 4.5 uses 15,954 output tokens per SWE-Bench Pro task on average, versus 67,020 for Opus 4.8 (max), a 4.2x difference. Source: xAI – Grok 4 5

CursorBench contamination note: Cursor's launch blog disclosed that "Grok 4.5 has an advantage on CursorBench because an earlier snapshot of the Cursor codebase was accidentally included in training. The exact impact is unclear. That data has been removed for future models." CursorBench scores were therefore excluded from their published comparison. Source: Cursor – Grok 4 5

Artificial Analysis independent evaluation

Source: Artificialanalysis – Grok 4 5

Notes:

• Grok 4.5 is ranked #13 out of 184 models on the Intelligence Index. The top model on the same leaderboard scores higher (Fable 5 at undisclosed exact score, but the chart shows it above Grok 4.5). Source: Artificialanalysis – Grok 4 5
• At 57.9 tokens/sec, Grok 4.5 is slower than average for its price tier (median 64.8). This contradicts xAI's claim of "80 TPS" in the announcement, which may refer to a different measurement methodology (e.g., non-reasoning mode or peak throughput). Source: Artificialanalysis – Grok 4 5 , xAI – Grok 4 5
• The Artificial Analysis page lists the cache hit price as $0.50 (pre-July 24 drop); the current live price is $0.30. Source: Twitter – Status

Security benchmark (Dam Secure, July 11)

An independent PR security review benchmark by Dam Secure tested 10 models on finding planted access-control bugs in pull requests. Source: Damsecure – Pr Review Security Benchmark

Grok 4.5 lands on the cost/quality frontier at $0.20 per PR (cheapest among frontier models), with F1 0.77. It is dominated by GPT-5.6 Sol on quality (F1 0.91) but at roughly one-third the cost per PR. Source: Damsecure – Pr Review Security Benchmark

Community build-off comparison (TryAI, July 8)

TryAI ran a one-shot build-off giving Grok 4.5, GPT-5.5, Claude Opus 4.8, and Claude Fable 5 the same three prompts to build interactive HTML apps. Results: Source: Tryai – Grok 4.5 Vs Gpt 5.5 Vs Claude Build Off

• 3D Rubik's Cube: Grok 4.5 failed on first attempt (blank void), succeeded on retry. Opus 4.8 and Fable 5 succeeded first try. Winner: Claude models.
• Particle gravity sandbox: All four succeeded. GPT-5.5 won on visual quality. Grok 4.5 was clean and orbital.
• Breakout game: All four produced ship-quality games on first try. Dead heat.
• Speed and cost: Grok 4.5 was fastest (first token in 0.44s, 110 tok/s throughput in their harness) and cheapest ($0.002/reply), roughly half the cost of GPT-5.5 and Opus 4.8. Source: Tryai – Grok 4.5 Vs Gpt 5.5 Vs Claude Build Off

Latest News

Grok 4.5 Launch (July 8-16, 2026)

Grok 4.5 launched July 8, 2026, with the HackerNews submission reaching 776 points and 1,502 comments. The x.ai/news/grok-4-5 page is dated July 16. The model is a mixture-of-experts architecture trained jointly with Cursor, on "tens of thousands of NVIDIA GB300 GPUs," with reinforcement learning on "hundreds of thousands of tasks, centered on multi-step software engineering." It is now the default model in Grok Build, available in Cursor on all plans, in the SpaceXAI console API, and in GitHub Copilot (added July 28). xAI claims 80 TPS serving speed and 2x token efficiency versus comparable leading models. Free usage is offered for a limited time in Grok Build and Cursor. Sources: xAI – Grok 4 5 , Cursor – Grok 4 5 , News – Item , GitHub – 2026 07 28 Grok 4 5 Is Now Available In Github Copilot

Grok Build Open-Sourced (July 15, 2026)

xAI published the full Rust source code for the Grok Build CLI at GitHub – Grok Build under Apache 2.0. The repo includes the agent runtime (xai-grok-shell), the TUI (xai-grok-pager), tool implementations (xai-grok-tools, with in-tree source ports of openai/codex and sst/opencode tool implementations), workspace management (xai-grok-workspace), and supporting crates for config, MCP, markdown, and sandboxing. It reached 23.6k stars and 4.5k forks within two weeks. However, the CONTRIBUTING.md explicitly states "External contributions are not accepted," and the repo is "synced periodically from the SpaceXAI monorepo" rather than developed in the open. The HackerNews thread reached 590 points and 644 comments. Sources: GitHub – Grok Build , News – Item

Grok Build CLI Data Exfiltration Scandal (July 12-14, 2026)

Independent researcher @cereblab published a reproducible wire-level analysis of Grok Build CLI 0.2.93 showing three findings: (1) file contents including .env secrets are transmitted unredacted to xAI via both the model-turn API and a storage upload; (2) the entire repository is uploaded as a git bundle including all tracked files and git history, independent of what the agent reads, with no storage size ceiling in tested range (5.10 GiB uploaded from a 12 GB repo, all HTTP 200); (3) the storage destination is the GCS bucket grok-code-session-traces, and disabling "Improve the model" does not stop the upload. xAI disabled the upload server-side after the report went public (server returned disable_codebase_upload: true and trace_upload_enabled: false), and Elon Musk publicly committed to deleting previously uploaded data. No security advisory, retention policy, or incident statement was published. A second researcher (@wetlink) independently confirmed the findings and published a privacy-hardening config tool. The HackerNews thread reached 539 points and 229 comments. Sources: Gist – Dc9A40Bc26120F4540E4E09B75Ffb547 , News – Item , Internationalcyberdigest – Xais Grok Build Cli Uploads Entire Git Repositories To A Google Cloud Bucket , GitHub – Grok Build Privacy Hardening

Grok 4.5 Cache Price Reduction (July 24, 2026)

The cached input price for Grok 4.5 dropped from $0.50 to $0.30 per MTok (under 200K), and from $1.00 to $0.60 (above 200K). This was announced via X/Twitter but no corresponding blog post or changelog entry was published. Source: Twitter – Status

Grok 4.5 in GitHub Copilot (July 28, 2026)

GitHub added Grok 4.5 as a selectable model in Copilot. The changelog entry does not specify whether it is available on all plan tiers or limited to specific ones. Source: GitHub – 2026 07 28 Grok 4 5 Is Now Available In Github Copilot

Community Signals

July community discussion around xAI was dominated by three events in sequence: the Grok 4.5 launch (July 8), the data exfiltration scandal (July 12), and the open-source release (July 15). Sentiment shifted from cautious optimism about the model to alarm about the CLI, to mixed reactions on the open-source release.

Grok 4.5 launch thread (776 points, 1,502 comments)

The largest HN thread on xAI in the tracked period. Sentiment ranged from genuine benchmark analysis to skepticism about the Cursor training data integration.

• A TryAI build-off comparison (173 points, 93 comments) concluded: "Grok 4.5 is the speed-and-value monster... It built a great Breakout and a slick gravity sim, streams about twice as fast as the field, and is the cheapest to run. Its only real miss was the hardest stateful task (a blank cube on the first try, fixed on the retry)." Source: Tryai – Grok 4.5 Vs Gpt 5.5 Vs Claude Build Off , News – Item
• A security benchmark by Dam Secure (11 points) found Grok 4.5 lands on the cost/quality frontier for PR security review at $0.20/PR with F1 0.77, but trails GPT-5.6 Sol (F1 0.91). Source: Damsecure – Pr Review Security Benchmark , News – Item
• Cursor's own launch blog disclosed that Grok 4.5 had an "advantage on CursorBench because an earlier snapshot of the Cursor codebase was accidentally included in training." Source: Cursor – Grok 4 5

Data exfiltration thread (539 points, 229 comments)

The second-largest xAI thread of the month. Sentiment was overwhelmingly critical, focusing on privacy, legal liability, and the contrast with how other coding agents handle codebase context.

• freakynit, reacting to the whole-repo upload finding: "It uploads the whole repository, every tracked file's content plus git history, independent of what the agent reads. Holy cow!!!! I mean I kinda expected Elon would do something like this to try to catch-up, but this is extremely concerning. This is precisely the reason, even though their pricing is competitive and grok-4.5 is actually good enough, I chose not to go with them." News – Item
• whateveracct, on the legal dimension: "it's straight up data exfiltration and should be illegal." News – Item
• Stromgren, on why this differs from normal coding agent behavior: "There's a hell of a difference between a tool that asks my permission to read a file to make it part of a prompt and a tool that packages up my whole working directory and sends it to Google Cloud Storage." News – Item
• gitgud, on the risk of proprietary coding agent runners: "This is one of the reasons why native proprietary coding agent runners like claude-code, codex, grok-build etc are so dangerous for privacy, you just don't know what 'secret sauce' they'll add in the next update. It's much safer to use something like opencode and use models via their API, however, the tradeoff is that it will never perform as well as it does in their native agent runners." News – Item
• gruez, with a counterpoint: "Give enough usage, you can reconstruct an entire codebase via tool calls alone, and it'll be entirely undetectable because it's all done server side. Whatever grok's doing is just more blatant, but using opencode or whatever doesn't create a meaningful security boundary." News – Item
• taywrobel (GitHub Copilot engineer), in a notable defense of GitHub's data practices: "Very few employees within GitHub proper even have access to view private repositories, and in the rare cases where that's done for legal or safety reasons the repository owner is notified. There are currently no OpenAI employees with access to GitHub systems, so there's about 4 layers of protection in place to prevent private repositories access." News – Item
• kordlessagain, providing the local mitigation: "export GROK_TELEMETRY_TRACE_UPLOAD=0, export GROK_TELEMETRY_ENABLED=0, or config file with [telemetry] trace_upload = false, [harness] disable_codebase_upload = true. The practical takeaway for users: your entire codebase leaves (uploaded) your machine unencrypted on each Grok Build invocation, not just files you ask it to read, and no visible setting stops it." News – Item
• danudey, on the severity: "Imagine the fallout if someone discovered a method of read access to this bucket." News – Item

Open-source release thread (590 points, 644 comments)

The third major xAI thread. Reactions were positive about transparency but critical of the no-contributions policy and the timing relative to the exfiltration scandal.

• The repo's own CONTRIBUTING.md: "External contributions are not accepted." Source: GitHub – Contributing.Md
• A blog post titled "Grok Build Is Open Source, but xAI Won't Take Your Patches" noted the contradiction. Source: Vincentschmalbach – Grok Build Open Source Contributions Closed , News – Item

Overall community sentiment

The dominant July themes are: (1) Grok 4.5 is viewed as genuinely competitive on benchmarks and speed, with the TryAI build-off and Dam Secure security benchmark providing third-party confirmation; (2) the data exfiltration scandal is treated as a disqualifying trust failure for enterprise use of the Grok Build CLI, even after the server-side fix, because the client retained the capability and the company published no advisory; (3) the open-source release is welcomed for transparency (the community can now audit the upload pipeline) but the no-contributions policy is seen as "source-available" rather than truly open source; and (4) the CursorBench training contamination disclosure raised questions about benchmark integrity that extend beyond xAI to the joint model's provenance.

Enterprise Readiness

The consumer pricing and business pages remain 403-blocked to automated fetch. Enterprise feature verification carries forward from May where noted. The data exfiltration scandal is a material new enterprise risk.

Terms explained:

• Data residency - the ability to specify which geographic region (US, EU) your API requests are processed in. Grok 4.5's absence from eu-west-1 means EU customers cannot keep traffic within the EU when using the flagship model. xAI – Grok 4.5
• Air-gapped / on-prem - the ability to run the coding agent with no network connectivity to the provider. Grok Build is now open source (the harness), but the model (Grok 4.5) is proprietary and only served from xAI's API. Local-first operation is not possible.

Transparency Gaps